Ranking in AI Answers

To appear in ChatGPT, Claude or Perplexity answers: be crawlable without JavaScript, answer the question directly in a quotable form, present as one consistent entity everywhere, and give engines specifics worth citing. Then measure it — ask the engines your buyers’ questions monthly and log who gets cited.

The practitioner’s sequence

1. Audit what the engines actually receive

Fetch your own pages the way a crawler does — no JavaScript execution — and read what returns. GPTBot fetches your bundle and never runs it. ClaudeBot downloads and does not execute. PerplexityBot parses static HTML.

Whatever is missing from that response is invisible, permanently, regardless of how good it is. This is the gate, and it is the single most common reason a business is absent from answers about its own category. Ten minutes to test, and the result is binary.

Fix this before anything else on this page. The technical article covers the plumbing in depth.

2. Restructure to answer-first

Engines extract passages. Self-contained passages win, and passages that require the preceding three paragraphs to make sense do not get lifted.

The pattern: each page opens with the complete answer in roughly sixty words, then earns the elaboration. Headings phrased as the questions buyers actually type become extraction handles — a heading that reads “Our Approach” gives an engine nothing, while “How does reconciliation actually work at scale?” gives it a labelled unit of meaning.

This is the highest-leverage writing change available, and it requires no new pages — only rewriting the openings of the ones that carry commercial weight.

3. Consolidate the entity

An engine composing an answer must resolve who is claiming this. That resolution is helped by consistency and destroyed by contradiction.

Same name, same description, same claims across your site, your LinkedIn, your GitHub, any directory that lists you. Organization, Person and Service schema in one connected graph with stable identifiers rather than islands of markup on individual pages.

Engines cross-check before citing. Two plausible readings of who you are means neither gets asserted confidently, and the citation goes to whoever was unambiguous.

4. Feed specifics

Engines lift mechanisms, numbers with sources, and named trade-offs far more readily than adjectives. “We deliver excellence” is uncitable by construction — there is nothing in it to quote. “India completed the move to T+1 equity settlement in January 2023 under SEBI’s phased rollout” is citable, checkable, and useful to an engine answering a settlement question.

This is why the fourteen teardowns on this site earn citations while a capabilities page would not: they say checkable things about real systems, and a checkable thing is the unit of currency here.

5. Measure with a panel

There is no rank tracker, so you build the instrument.

Write the questions your buyers actually ask — twenty to thirty of them, in their words rather than as keywords. If you cannot write thirty, interview three customers and you will have forty.

Run them monthly against each engine from a clean session, logging three things: whether you were cited, who else was, and what the answer said about your category. That third column is quiet competitive intelligence you cannot obtain any other way.

Read the trend rather than the month. Citations are noisy question by question; what matters is whether your share of the panel rises across quarters.

The uncomfortable truths

Authority accrues over months. There is no AI-search equivalent of buying ads. The entities engines cite confidently are the ones with a consistent, corroborated presence over time, and that cannot be shortcut.

Engines change without changelogs. Behaviour shifts, sometimes noticeably, with no announcement and no documentation. Build to be citable in general rather than tuned to one engine’s current preferences — the general version survives; the tuned version becomes technical debt.

You will be quoted verbatim, without context. Passages get extracted without your framing and without the paragraph that softened the claim. Content that would embarrass you when quoted alone will, eventually, be quoted alone. That is a stricter writing standard than marketing copy usually meets, and it is the standard this entire site is written to.

Nobody can guarantee you become the answer. Engines blend sources and weight signals you do not control. What is buyable is citability; what is not buyable is a position. Any vendor claiming otherwise — including any version of me that ever did — should be disbelieved.

Why this is converging with ordinary quality

The AI-citation bar and the survives-a-Google-update bar are becoming the same bar. Both reward content that answers a real question specifically, cites sources, and could be read aloud to an expert without embarrassment. Both punish thin pages, keyword scaffolding and unsupported claims.

That convergence is genuinely good news for planning: you do not need a separate AI-visibility programme. Build a site that is crawlable without JavaScript, structured answer-first, consistent about its identity and dense with checkable specifics — then measure both surfaces and stop worrying about which one you are optimising for.

What a citable page looks like structurally

Take any page on this site and the shape repeats, deliberately.

A sixty-word opening that fully answers the page’s question, written so it survives being lifted with no surrounding context. If an engine quoted only that paragraph, the reader would still have received something true and useful.

Headings as questions, so each section is a labelled unit an engine can extract independently.

Specifics with sources named inline — dates, regulations, thresholds, mechanisms. Not linked footnotes, which are stripped on extraction, but the source named in the sentence itself so it travels with the claim.

Explicit limits. The section on every service page saying what it will not do is not modesty; it is citable content. Engines answering “can anyone guarantee AI rankings” need a source that says no clearly, and pages that hedge everything provide nothing quotable.

An FAQ block in schema, because question-and-answer pairs are the format closest to what an engine is assembling.

None of that is exotic. It is ordinary careful writing with the extraction step kept in mind — and the useful property is that a page built this way also reads better to humans, which is why the two bars keep converging.

Where to start on Monday

Three days of work, in order, covering most of the available ground.

Day one: the gate. Fetch your twenty most commercially important pages without JavaScript. Fix whatever is missing before touching anything else, because every later improvement is invisible until this passes.

Day two: the openings. Rewrite the first paragraph of those twenty pages so each states its answer completely, in plain prose, with no preamble. Convert the headings into buyer questions while you are there.

Day three: identity and proof. One connected schema graph with stable identifiers, consistent naming across every external profile you control, and a pass through the claims on those pages attaching a source or deleting the claim.

After that it is content depth and time, neither of which can be rushed. But those three days remove the structural reasons you are currently being skipped, and they cost nothing but attention.

Related reading

/blog/generative-engine-opt — the framework in full · /blog/technical-seo — the rendering gate · /blog/google-updates — why the two bars converge · /services/generative-engine-opt — the work, done for you